Search

Kimi K3 Pricing Ends the 'Open Weights = Cheap' Era

Moonshot AI released Kimi K3 on July 16, 2026 with API pricing that directly matches Anthropic's Claude Sonnet 5 tier: $3.00 per million input tokens and $15.00 per million output tokens. This is not a mistake. It is a deliberate signal that "open weights" no longer means "cheap".

The old rule was simple. Chinese open-weight models undercut US closed APIs on price. DeepSeek V4 Pro costs $0.435 input / $0.87 output. Qwen3.7 Max costs even less. Enterprises picked open weights to save money and accepted slightly lower capability. Kimi K3 breaks that tradeoff. It sells at Sonnet 5's price point, claiming frontier capability at frontier pricing. The open-weights category now spans both premium and budget, and buyers must re-evaluate what they are actually paying for.

Kimi K3: Specs and Price Structure

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 896 experts and a 1-million-token context window, according to the Kimi API platform. The model weights are scheduled for public release on July 27, 2026, following a short embargo period.

The API pricing is straightforward:

  • Standard input: $3.00 / MTok
  • Standard output: $15.00 / MTok
  • Cached input: $0.30 / MTok (via Mooncake split-inference architecture claiming >90% cache-hit rates for coding workloads)

This represents a 3.2x increase for input and 3.75x increase for output compared to Moonshot's previous K2 generation. Moonshot is not competing on price anymore. It is competing on capability and branding, targeting the premium frontier segment that Anthropic and OpenAI currently dominate.

How Kimi K3 Stacks Up Against US Frontier Models

Priced at $3/$15, Kimi K3 sits directly alongside Claude Sonnet 5. However, Anthropic is running introductory launch pricing of $2/$10 through August 31, 2026, making K3 currently 50% more expensive than Sonnet 5's promo rate. Buyers comparing today should account for that discount window.

Against the absolute flagships, K3 costs less. Claude Fable 5 runs $10/$50. GPT-5.6 Sol runs $5/$30. K3 offers a lower-cost entry to frontier reasoning while retaining the open-weights benefit. But the open-weights benefit comes with caveats.

The Open-Weight Reality Check

The license drops on July 27, but self-hosting a 2.8T-parameter MoE model requires massive enterprise GPU infrastructure. Estimates point to 64+ accelerators minimum. For most developers, self-hosting is not viable. They will use the API regardless of the open license, and they will pay premium API prices.

This is the key tension. The open-weights label carries expectations of low-cost inference and deployment flexibility. Kimi K3 delivers neither at the API level. The open license exists in legal terms but not in practical economic terms for anyone outside large enterprises with dedicated clusters.

Moonshot argues that the effective input cost is $0.30 via high cache-hit rates, making the real cost comparable to cheaper models. That holds for workloads with repetitive context patterns. For heavy reasoning workloads generating long output traces, the $15/MTok output charge remains a significant margin risk. Cache hits do not reduce output cost.

Kimi K3 vs. DeepSeek V4 Pro: The New Open-Weight Divide

DeepSeek's V4 Pro (1.6T parameters, released April 2026) costs $0.435 / MTok input and $0.87 / MTok output. That is 7x cheaper for input and 17x cheaper for output than Kimi K3. DeepSeek V4 Pro is the budget open-weight champion. Kimi K3 is the premium open-weight challenger.

Both are open weights. Both are Chinese lab models. Both score near the frontier on benchmarks. But their pricing strategies could not be more different. This bifurcation means the category "open-weight models" no longer signals anything about cost. Buyers must evaluate each model individually rather than using a blanket heuristic.

What This Means for Enterprise Buyers

The decision matrix for choosing between closed US APIs and open Chinese weights has changed. Three factors now govern the choice:

  • Effective cost per task: Kimi K3's verbosity in reasoning traces increases its effective cost-per-task to approximately $0.94, according to a 2026 cost analysis, while cheaper open models like DeepSeek complete similar tasks for a fraction of that. Cache-hit rates matter, but output-heavy workloads erase the advantage.
  • Self-hosting feasibility: K3's 2.8T parameters require enterprise infrastructure. Most teams cannot self-host. The open license is irrelevant if you cannot run it.
  • Capability gap: If Kimi K3 genuinely matches Sonnet 5 on coding and reasoning benchmarks, enterprises pay the same price for comparable capability with the optionality of open weights. If the gap is smaller than claimed, they overpay.

Enterprise buyers should treat Kimi K3 as a premium API product that happens to have an open license you probably cannot use. Compare it against Claude Sonnet 5 on your specific workloads, measure effective cost including cache hits and output length, and ignore the open-weights label until you verify that self-hosting is actually viable for your infrastructure.

See also: Best Budget Models · Top Overall Models · Best Open-Weight Models

The open-weights category now includes both DeepSeek V4 Pro at $0.435/$0.87 and Kimi K3 at $3/$15. The word "open" does not mean "cheap". It never did. But until now, the market could pretend otherwise. Kimi K3 ends that pretense.